DIPNet: Driver intention prediction for a safe takeover transition in autonomous vehicles

Bonyani, Mahdi; Rahmanian, Mina; Jahangard, Simindokht; Rezaei, Mahdi · 2023 · Crossref

DOI: 10.1049/itr2.12370

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety challenge of predicting driver intention to facilitate safe takeover transitions in Level 3 autonomous vehicles. As automated systems increasingly handle driving tasks, the ability to anticipate when a human driver intends to intervene—such as during lane changes or turns—is essential for preventing accidents and ensuring smooth control handovers. The authors identify limitations in existing methods, particularly the reliance on manual feature extraction and the neglect of the small, unbalanced nature of available datasets, which leads to overfitting. To resolve these issues, the study proposes DIPNet, a novel deep neural network framework designed to predict driver maneuvers up to 4 seconds in advance using only video data, without requiring handcrafted features or external sensor inputs. The methodology employs a four-stream deep learning architecture that processes both in-cabin (driver observation) and out-cabin (road view) video data. The model integrates DenseNet for spatial feature extraction, Long Short-Term Memory (LSTM) networks for temporal feature analysis, and optical flow models (FlowNet2 and RAFT) to capture motion dynamics. A global attention mechanism is incorporated to prioritize significant features, while an implicit semantic data augmentation algorithm is used to mitigate dataset imbalance and improve generalization. The system was trained and evaluated using the public Brain4Cars dataset, which contains synchronized video pairs from 10 drivers across five maneuver classes: going straight, left/right lane changes, and left/right turns. The dataset was split into training, validation, and testing sets, with performance assessed using accuracy, precision, recall, and F1-score metrics via K-fold cross-validation. The experimental results demonstrate that DIPNet outperforms state-of-the-art models in the majority of driving scenarios. By leveraging an end-to-end data-driven approach, the model effectively captures complex spatial and temporal dependencies from video inputs alone. The integration of optical flow and attention mechanisms allows the system to distinguish between subtle driver behaviors and environmental cues, achieving higher prediction accuracy compared to previous methods that relied on manual feature engineering or single-modality data. The study confirms that combining in-cabin and out-cabin video streams provides complementary information that significantly enhances prediction performance, even with the limited size of the Brain4Cars dataset. The significance of this work lies in its contribution to the development of safer, more responsive autonomous driving systems. By providing a reliable, real-time prediction of driver intent several seconds before a maneuver occurs, DIPNet addresses a key prerequisite for Level 3 automation, where the system must seamlessly transition control to the human driver. The proposed framework’s low computational complexity and high accuracy make it suitable for practical deployment in vehicles. Furthermore, the study highlights the effectiveness of data-driven approaches over traditional model-based methods, offering a scalable solution that can adapt to diverse driving conditions and sensor variations without extensive manual tuning.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms
• anticipation
• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software
• Theoretical Contribution: computational model, conceptual framework